Real-time Aftershock Forecasting in Turkey (RAFT)

Trapped between the Eurasian tectonic plate and the northward moving Arabian plate, the Anatolian plate is being squeezed to the west forming two major earthquake fault zones, the North and East Anatolian Fault Zones. Throughout recorded history both have generated fatal earthquakes. The city of Antakya, for example, was destroyed twice in 115 and in 556 probably killing more than 500000 people; to the north the 1939 Erzincan earthquake killed more than 30,000 people and more recently the Izmit and Duzce earthquakes together killed more than 18000. In the last century Turkey has suffered 43 fatal earthquakes killing in the region of 100,000 people and, on average, we can expect a killer earthquake every 2 years.

While it is not possible to predict earthquakes, very robust statistical laws in seismology and the rapid calculation of complex stress fields produced by large earthquakes, has allowed us to make dependable, physics-based forecasts, resulting from detailed, near real-time seismological observations, of where aftershocks are likely to occur.

In this project we plan to continue to develop this science so that emergency managers can include science-based forecasts in operational decision-making during an earthquake crisis.

Statistical forecast methods depend on accurate, plentiful data on the location and sizes of the developing aftershocks. Normal seismic infrastructure is only capable resolving larger events, reducing the power and stability of the statistics. We will purchase and test a custom designed completely dedicated portable seismic network which, when combined with the existing Turkish seismological network, will be capable of recording earthquakes of M >2. This system will be equipped with telemetry so that the seismic catalogues are available at an Aftershock Management Centre in near real-time.

While the statistical techniques have already been shown to produce good forecasts of the probability of aftershocks, additional work will allow the automation of many of the decisions which must be made concerning, for example, when to move to a more complex aftershock model during a crisis. Such subjective decisions will be extremely important in making sure that physics, not emotion, guide the decision-making process, which will, of course, ultimately be led by a trained seismologist.

During the recent Nepal earthquake crisis the University of Ulster group provided near real-time forecasts both to government and to a major NGO which included earthquake space-time density maps as well as underlying stress distributions to provide qualitative forecasts of the spatial structure of future events. This work will be developed to produce more quantitative spatial estimates of aftershock probability, based on the evolving theories of rate and state friction, which again, can be automated to provide subjective risk evaluations.

So that the entire process can be rigorously tested, we will deploy the network and test every aspect of the protocols which evolve during the project, following a moderate earthquake occurring in the last 6 months of the project. This field trial will be observed by an International Advisory Panel of scientists and humanitarian workers which will be constituted early in the project.

Given that this is an explicitly applied project most of the benefit will be in terms of the impact on the management of the aftershocks to large destructive earthquakes both in Turkey and worldwide.

There are however significant academic and research advances which have already arisen from our work on aftershock forecasting and which will be completed and published early in the course of the new project. In particular, the use of subjective decision-making in the choice of events which qualify as significant in that the change the complexity of the model required best to describe the data, has important academic spin-offs, which will be of interest to a wide swathe of earthquake scientists. One of the problems that has been addressed by many in the literature of statistical seismology is the definition of earthquakes which are causally clustered so that we might examine the so-called background seismicity. While not being the ambition of our research, it turns out that the Bayesian Monte Carlo techniques used for the objective modelling of data uses the Bayesian information criterion to identify "significant" earthquakes within the sequence. These events effectively spawn new aftershock sequences and in some real sense they are more significant than the other events in the sequence. We are still working on the detail of this idea, but it's clear from our work so far that these "significant" earthquakes depend for their definition on the magnitude of completeness of the catalogue. Thus the details of the de-clustering of seismic catalogues may be an artificial function of the seismic network which recorded them.

Our objective fitting of an increasingly complex model using the Bayesian information criterion to decide when extra parameters are justified by the improved fit to data may also be generally interesting. In particular our use of Bayesian Monte Carlo modelling coupled with objective judgements adjudicated by an appropriate information criterion allows us not only to define the most parsimonious model which adequately describes the data but also allows us to track the uncertainties in our parameter estimates in near real-time as the data stream is analyzed.